Week 9 Emily Hand UNR
This Week Blended Template Partial SVM Model Speed Motion Model Improvements/Debugging/Testing Finding best parameters Partial SVM Model Debugging/Testing Speed Decrease Speed of both methods Motion Model Used to help detector
Blended Template SVM for entire template is used to find the location in the next frame Confidence of each block extracted using blockSVM function A new template is constructed from positive blocks Blended with most recent positive blocks to form a complete template
Blended Template Ex 1
Example 1 Tracked Successfully: 36 Frames First frame – full occlusion 40% Partial Occlusion: Frames 33-36 Full Occlusion: Frame 37 Time Per Frame: ~5 minutes Threshold: 0.40
Example 1 Partial
Example 2
Example 2 Tracked Successfully: 57 frames First full instance – end of sequence 100% Partial Occlusion: Frames 40-48 Full Occlusion: None Time Per Frame: ~5 minutes Threshold: 0.40
Example 2 Partial
Example 3
Example 3 Tracked Successfully: 21 Frames First full instance – full occlusion 62% Partial Occlusion: Frames 66-69, 71-74 Full Occlusion: Frame 76 Time Per Frame: ~6 minutes Threshold: 0.40
Example 3 Partial
Example 3 Partial
Example 4
Example 4 Tracked Successfully: 84 Frames First full instance – end of sequence 100% Partial Occlusion: Frames 35-42 Full Occlusion: None Time Per Frame: ~7 minutes Threshold: 0.40
Example 4
Example 5
Example 5 Tracked Successfully: 133 Frames First full instance – end of sequence 100% Partial Occlusion: Frames 180-185, 199-203 Full Occlusion: None Time Per Frame: ~30 seconds Threshold: 0.40
Example 5 Partial
Example 5 Partial
Example 6
Example 6 Tracked Successfully: 46 Frames First full instance – full occlusion 56% Partial Occlusion: Frames 30-32, 42-46 Full Occlusion: 47-50 Time Per Frame: ~30 seconds Threshold: 0.40
Example 6 Partial
Example 6 Partial
Example 7
Example 7 Tracked Successfully: 80 Frames First full instance – end of sequence 100% Partial Occlusion: Frames 147-155 Full Occlusion: None Time Per Frame: ~30 seconds Threshold: 0.40
Example 7 Partial
Conclusions Blended Template Model works for situations without full occlusion. In the case of full occlusion a motion model must be used This can improve our results and allow the detector to continue locating the person after the full occlusion has passed.
Partial SVM Test with entire template. Find best match. If low confidence, then find blocks with a high confidence and train new SVM with those blocks. In next frame, test with both the entire template and the partial SVM model. Find highest match between the two. Worse Results than we expected....
Partial SVM
Partial SVM
Motion Model Used to help detector. A predicted location is computed. The confidence of each position inside of the search area is computed. The distance of each of these points is computed from the predicted location. If the distance is high, then the confidence at that point is decreased.
Blended Template Ex 1 36 Frames → 82 40% → 90%
Example 6 46 Frames → 82 56% → 100%
Speed Improvement Speed Decreased Search Area 7 minutes → 30-45 seconds per frame Much faster If the confidence is low for several frames, then the search area grows Once the confidence is high again, the search area shrinks around the confident area.
What's Next? New detection ideas 10-15 blocks in a template. Each has its own SVM Search for these parts and combine them to find best detection.